Web and Social Network Analytics

Week 5: Ethics in Social Network Analytics

Dr. Zexun Chen
Zexun.Chen@ed.ac.uk

Feb-2025

Table of Contents

GDPR

Question 1

What ethical responsibility do we have when gathering and using information, behavior patterns, and recommendation systems?

Inequalities

  • What went wrong?
  • What can we do better next time?

General Data Protection Regulation (GDPR)

The General Data Protection Regulation (GDPR) is an EU law governing data protection and privacy. It is a key component of EU privacy law and human rights law (Article 8, Charter of Fundamental Rights of the European Union).

Note: The UK now follows UK GDPR.

GDPR Keypoints

  • Privacy
  • Fairness
  • Transparency, accountability, explainability

Privacy

🔍 Anonymized data can never be fully anonymous!

  • De Montjoye et al. (2013)
    • Finding: Four spatio-temporal data points can uniquely identify 95% of individuals.
  • Li et al. (2016)
    • Finding: Over 50% success rate in inferring age, gender, education from location data in geo-social networks.

Privacy-Preserving Techniques

✅ Differential Privacy (DP)
- Adds mathematical noise to protect individual data points.
- Example: Used by Apple & Google in analytics.

✅ Federated Learning (FL)
- Decentralized AI training → No raw data is shared.
- Example: Google’s Gboard keyboard learns from users locally.

✅ Homomorphic Encryption (HE)
- AI can process encrypted data without decrypting it.
- Example: Used in healthcare AI for private medical analysis.

✅ Synthetic Data
- AI-generated artificial data mimicking real datasets.
- Example: Amazon & NVIDIA use it to train AI without privacy risks.

✅ Secure Multi-Party Computation (SMPC)
- Multiple parties collaborate to compute results without exposing private data.
- Example: Used in financial risk analysis.

Question 2: Trade-offs in Privacy & Convenience

What trade-offs are you willing to accept between convenience and privacy?

Examples of personal data sharing:

  • Purchase history → Amazon recommendations
  • Health & Activity → Extra days off for Fitbit performance
  • Economic status → Airfare pricing based on affordability
  • Banking history → Discounts & recommendations
  • Glassdoor → Access to salary info in exchange for sharing your own
  • Private conversations → Facebook Messenger influencing Spotify recommendations

Data Sharing Willingness

Fairness in AI: Everything is Fine?

🔍 Case Study:

Amazon shut down its AI hiring tool after discovering gender bias, as the system penalized female candidates.

Bias in AI Decision-Making

📌 ProPublica Investigation:
A recidivism risk scoring tool was biased against African Americans, leading to unjust sentencing disparities.

Algorithmic Bias: Where Does It Come From?

Bias comes from: \[ Data + Model/Algorithm = Prediction (Decision) \]

Common bias sources:

  • Data bias
  • Algorithmic bias

What is Data Bias?

📌 Data Bias occurs when:

  • Missing key attributes that influence predictions.
  • Human-generated content contains biases.

Fact:
Most datasets are biased unless generated from carefully controlled randomized experiments.

Types of Data Biases

  • Response or Activity Bias:
    • Biases from reviews, social media posts, Wikipedia edits, etc.
  • Selection Bias due to Feedback Loops:
    • Ads, content personalization, recommendations reinforce bias.
  • Bias due to System Drift:
    • Data generation processes change over time.
  • Omitted Variable Bias:
    • Missing critical attributes affecting outcomes.
  • Societal Bias:
    • Social media & web content reflect human biases.

Can We Simply Ignore Biased Attributes in Modeling?

❌ NO!

  • Ignoring sensitive attributes like gender & race does not remove bias.
  • Such biases have already been embededing in other attributes

How to Handle Data Bias?

⚖️ Bias Mitigation Techniques fall into three categories:

✅ Pre-processing (before training) → Modify data to reduce bias
✅ In-processing (during training) → Adjust models to reduce discrimination
✅ Post-processing (after training) → Adjust predictions to ensure fairness

Debiasing Techniques & Tools

Pre-processing Approaches (Fix biased data before training)

🔹 Re-sampling: Adjust class distributions (SMOTE for imbalanced datasets)
🔹 Re-weighting: Assign weights to samples to balance representation
🔹 Fair Representation Learning: LFR (Learning Fair Representations)

🛠 Frameworks:
✅ AI Fairness 360

✅ Fairlearn

In-processing Approaches (Modify learning algorithms)

🔹 Adversarial Debiasing: Train a second model to remove bias signals
🔹 Fairness Constraints: Use equalized odds, demographic parity
🔹 Differentially Private Training: Protects individual data points

🛠 Frameworks:
✅ Fairness Indicators (TensorFlow Extended)
✅ AIF360 Adversarial Debiasing Models

Post-processing Approaches (Correct biased outputs)

🔹 Equalized Odds Post-processing: Adjusts predictions for equal fairness across groups (More Info)
🔹 Calibrated Equalized Odds: Ensures fairness without sacrificing accuracy
🔹 Reject Option-Based Fairness: Tweaks uncertain predictions (More Info)

🛠 Frameworks:
✅ Fairlearn– Post-processing Tools
✅ AIF 360– Post-processing Fairness Algorithms

EU AI Act

EU AI Act: The First AI Regulation

The EU AI Act is the world’s first legal framework regulating AI, focusing on risk-based classification and accountability.

Key Provisions of the EU AI Act

  • Risk-Based Classification:
    • Unacceptable Risk 🚨 → Banned AI (e.g., mass surveillance, social scoring).
    • High Risk ⚠️ → Strict regulations (e.g., AI in hiring, healthcare, law enforcement).
    • Limited Risk ⚖️ → Transparency required (e.g., AI-generated content).
    • Minimal Risk ✅ → No restrictions (e.g., AI chatbots, video games).
  • Accountability & Transparency:
    • AI providers must ensure compliance with fairness, privacy, and transparency.
    • Strict penalties for non-compliance (up to €35M or 7% of annual turnover).

How Does the EU AI Act Impact AI Development?

  • AI companies must conduct risk assessments before deployment.
  • Increased regulatory oversight over high-risk AI applications.
  • Stronger protection for individuals against biased AI systems.

EU AI Act vs GDPR: What’s the Difference?

✅ GDPR → Protects personal data & privacy
✅ EU AI Act → Regulates AI models & their risks

Conclusion: Ethics & AI

  • Ethical AI & fairness should be a core consideration in social network analysis.
  • Transparency, accountability, and explainability are crucial for trustworthy AI.
  • Techniques for mitigating bias exist but require intentional application.